Application of Genetic Algorithms to Combinatorial Synthesis: A

Jasbir Singh,*,‡ Mark A. Ator,§ Edward P. Jaeger,| Martin P. Allen,‡. David A. Whipple,‡ James E. Soloweij,§ Swapan Chowdhary,. ⊥ and. Adi M...
0 downloads 0 Views 450KB Size
J. Am. Chem. Soc. 1996, 118, 1669-1676

1669

Application of Genetic Algorithms to Combinatorial Synthesis: A Computational Approach to Lead Identification and Lead Optimization†,∇ Jasbir Singh,*,‡ Mark A. Ator,§ Edward P. Jaeger,| Martin P. Allen,‡ David A. Whipple,‡ James E. Soloweij,§ Swapan Chowdhary,⊥ and Adi M. Treasurywala| Contribution from the Departments of Medicinal Chemistry, Biophysical and Computational Chemistry, Enzyme and Receptor Biochemistry, and Analytical Sciences, Sanofi Winthrop, 1250 South CollegeVille Road, CollegeVille, PennsylVania 19426-0900 ReceiVed September 15, 1995X

Abstract: A genetic algorithms (GA) based strategy is described for the identification or optimization of active leads. This approach does not require the synthesis and evaluation of huge libraries. Instead it involves iterative generations of smaller sample sets, which are assayed, and the “experimentally” determined biological response is used as an input for GA to rapidly find better leads. The GA described here has been applied to the identification of potent and selective stromelysin substrates from a combinatorial-based population of 206 or 64 000 000 possible hexapeptides. Using GA, we have synthesized less then 300 unique immobilized peptides in a total of five generations to achieve this end. The results show that each successive generation provided better and unique substrates. An additional strategy of utilizing the knowledge gained in each generation in a spin-off SAR activity is described here. Sequences from the first generations were evaluated for stromelysin and collagenase activity to identify stromelysinselective substrates. GlyProSerThr-TyrThr with Tyr as the P1′ residue is such an example. A number of peptides replacing Tyr with unusual monomers were synthesized and evaluated as stromelysin substrates. This led to the identification of Ser(OBn) as the best and most selective P1′ residue for stromelysin.

Introduction Recently, combinatorial/multiple synthesis of both oligomeric and non-oligomeric libraries of diverse compounds and highthroughput screening have provided a format for the identification of new lead compounds for various molecular targets.1 However, in any given template, the number of possible compounds one can synthesize in combinatorial or permutational2 libraries is enormous, often in the millions.1 Typically, † The work described here was carried out at Sterling Winthrop Pharmaceutical Research Division, before its divestiture on October 3, 1994, by Eastman Kodak to Sanofi Winthrop. Current Addresses: A.M.T., Allelix Biopharmaceuticals, 6850 Goreway Drive, Mississauga, Ontario L4V 1V7, Canada. E.P.J., 3 Dimension Pharmaceuticals Inc., 665 Stockton Drive, Exton, PA 19341. M.P.A. and D.A.W., Pfizer Central Research, Eastern Point Road, Groton, CT 06340. J.E.S., Amgen Inc., 1840 DeHavilland Dr., Thousand Oaks, CA 91320. M.A.A., Cephalon, Inc., 145 Brandywine Parkway, West Chester, PA 19380. S.C., Sanofi Winthrop, 31 Great Valley Parkway, Malvern, PA 19355. ‡ Department of Medicinal Chemistry. § Department of Enzyme and Receptor Biochemistry. | Department of Biophysical and Computational Chemistry. ⊥ Department of Analytical Sciences. * Author to whom correspondence should be addressed at NYCOMED Inc., 466 Devon Park Drive, P.O. Box 6630, Wayne, PA 19087-8630. ∇ Abbreviations: FMOC, (9-fluorenylmethoxy)carbonyl, CPG, controlled pore glass, AMP, ((aminopropyl)silyl)oxy, β-Ala, β-alanine, Acp, 6-aminocaproic acid, COP, 7-hydroxycoumarin-4-propionic acid, HOBt, Nhydroxybenzotriazole, mCl-t, a recombinant form of the human fibroblast collagenase, mSl-t, a recombinant form of the human fibroblast stromelysin. Single and three letter codes for 20 amino acids used for this work are as follows: A (Ala), D (Asp), E (Glu), F (Phe), G (Gly), H (His), I (Ile), K (Lys), L (Leu), M (Met), N (Asn), P (Pro), Q (Gln), R (Arg), S (Ser), T (Thr), U (denotes S-methylcysteine, Smc), V (Val), W (Trp), Y (Tyr). X Abstract published in AdVance ACS Abstracts, February 1, 1996. (1) For an excellent review see: (a) Gallop, M. A.; Barrett, R. W.; Dower, W. J.; Fodor, S. P. A.; Gordon, E. M. J. Med. Chem. 1994, 37, 12331251. (b) Gordon, E. M.; Barrett, R. W.; Dower, W. J.; Fodor, S. P. A.; Gallop, M. A. J. Med. Chem. 1994, 37, 1385-1401 and references cited therein.

0002-7863/96/1518-1669$12.00/0

one prepares libraries containing 104-106 compounds per template, assays these in a number of diverse screens, and subsequently identifies activess“hits”. High-throughput screening for a large number of targets is essential for success with this format. There have been a few attempts to evaluate the molecular diversity contained in a given library prior to synthesis.3 The motivation of this diversity assessment approach is to select a subpopulation which maximizes dissimilarity4 among the selected members. The synthesis and biological evaluation of this subpopulation rather then the entire library is a more manageable task. One of the important goals in combinatorial/multiple synthesis is to build high-fidelity libraries, which ensures greater probability of obtaining hits for a giVen biological target.5 An approach to accomplish this objective would be to synthesize diverse compounds in smaller sets (say, in the low hundreds at a time) and utilize the biological response to guide the selection of compounds6 for successive synthesis and biological evaluation (Scheme 1). Recently, genetic algorithms have been successfully utilized to find solutions to a number of complex problems.7-9 In fact, genetic algorithms (GA) are distinguished for their powerful optimization characteristics, enabling them to find a set of very good (but not necessarily the best) solutions rapidly where an (2) Pirrung, M. C. Chemtracts: Org. Chem. 1994, 7, 184-186. (3) Martin, E. J.; Blaney, J. M.; Siani, M. A.; Spellmeyer, D. C.; Wong, A. K.; Moose, W. H. J. Med. Chem. 1995, 38, 1431-1436. (4) See ref 3 for indices used to represent dissimilarity. (5) There is tremendous effort in producing and screening chemically diverse compound libraries. However, the real interest for the pharmaceutical industry is the biological diversity that is embodied in it, i.e. biodiversity not necessarily chemodiversity. (6) This approach would not require explicit description of a set of indices, but instead, the selection of indices would be implicit in this type of optimization approach. (7) For general references for genetic algorithms see: (a) Holland, J. H. Sci. Am. 1992, 66. (b) Forrest, S. Science 1993, 261, 872-878.

© 1996 American Chemical Society

1670 J. Am. Chem. Soc., Vol. 118, No. 7, 1996 Scheme 1. Biology-Guided Lead Identification Paradigma

Singh et al. Scheme 2. Summary of Variables Used for Genetic Algorithms

a Arrows represent input of biological data for generation Gen(i) to guide selection of compounds for generation Gen(i+1) (see text for details).

astronomically larger number of potential possibilities exists. Even though there have been numerous10 applications of GA, to the best of our knowledge, there is no example of an application of GA to guide chemical synthesis for structure optimization for any class of compounds. In this paper, we report the first application of GA-guided chemical synthesis. Our principle criteria for the choice of a test case to explore the usefulness of GA to guide chemical synthesis was to choose a template for which chemical synthesis has been well established, so that we could evaluate GA’s impact to find “hits” without confounding it with synthesis-related issues. Therefore, a peptide-based template was our first logical choice. Recently, we have reported the screening of immobilized peptide libraries as a tool for the determination of substrate specificity and selectivity for proteases.11 We reasoned that selection (and optimization) of hexapeptides consisting of 20 amino acids,12 representing 206 (64 000 000) possible structures, possessed all of the essential ingredients to be a good initial area in which to test the concept. The number of possibilities in the entire library was very large. The synthesis of these peptides had been previously worked out.11 A validated assay was in hand and available.11 Methods Genetic Algorithms. In this section, we will describe the basic ideas of the GA method, some issues involved in its use as a tool for the selection and representation of chemical structures, and finally the details of our implementation. GA optimization methods are based on several strategies from Darwinian theories of evolution. In the normal survival and evolution of the species, new genetic mutants constantly arise and their survival and “dominance” is based on their ability to find food, reproduce, and resist “assault” on their existence. These would be classified in the language of genetic algorithms as the “objective” function which is being optimized. In the same way, if our living mutating population was made up of hexapeptides instead of species of organisms and the evolutionary pressure being applied (the biological function) was the biological activity, then one could envisage exactly the same process occurring. The GA we use here is based on three basic strategies:13 selection, crossoVer, and mutation. The first of these strategies, selection, is the use of a breeding population in which the individuals (8) (a) Wagener, M.; Gasteiger, J. Angew. Chem., Int. Ed. Engl. 1994, 33, 1189-92. (b) Walters, D. E.; Hinds, R. M. J. Med. Chem. 1994, 37, 2527-2536. (c) Wehrens, R.; Lucasius, C.; Buyden, L.; Kateman, G. Anal. Chim. Acta 1993, 277, 313-324. (d) See ref 9 in ref 6a (listed above) for use of genetic algorithms for jet engine design. (9) (a) Judson, R. S.; Jaeger, E. P.; Treasurywala, A. M.; Peterson, M. L. J. Comput. Chem. 1993, 14, 1407-1414. (b) Judson, R. S.; Jaeger, E. P.; Treasurywala, A. M. J. Mol. Struct. (THEOCHEM) 1994, 308, 191206. (10) See refs 10-16 in ref 8a above. (11) Singh, J.; Allen, M. A.; Ator, M. A.; Gainor, J. A.; Whipple, D. A.; Soloweij, J. E.; Treasurywala, A. M.; Morgan, B. A.; Gordon, T. D.; Upson, D. A. J. Med. Chem. 1995, 38, 217-219. (12) We use all 20 coded amino acids (see a complete list under Abbreviations), except Cys. We employ S-methylcysteine (Smc, denoted by the single letter code U) as the 20th amino acid.

who are more “fit” in some sense (higher biological response in this application) have a higher chance of producing offspring and passing on their “genetic” information. The second strategy is the use of crossoVer14 (mating) in which a child’s genetic material is a mixture of his or her parents’. The final strategy is that of mutation, where the genetic material is occasionally “corrupted” to maintain a certain level of spontaneous and random genetic mutation in the population. The GA paradigm used here employed a modified version of the Genesis GA15 code and is outlined above in Scheme 2. We work with a population of individuals which interact through their genetic operators to carry out an optimization process. An individual is specified by a chromosome, a bit string in this case. Let us assume that a hexapeptide is to be represented by a bit string (i.e., a sequence of 1’s and 0’s) of 30 bits (or digits). Each amino acid is then represented by five bits: the first amino acid being coded into bits 1-5, the second being coded into 2-10, etc. Each five-bit code can essentially code for 25 or 30 unique amino acids. Since there are only 20 amino acids, this five-bit codon can easily accommodate a unique pattern for each amino acid.16 Therefore, the 30-bit string can be translated into a unique hexapeptide (and vice versa). A fitness function, also called the objective function (see above), is used to rank the individual’s chromosome. The optimization proceeds because the population produces individuals that have increasingly higher fitness. Initially, a set of Npop individuals is formed by choosing a set of Nb-bit strings at random and each member is synthesized and evaluated for fitness. A roulette wheel is conceptually created where the “slice” on the wheel for any given individual is proportional to the value, for that individual, for its fitness. Biologically more active peptides in our implementation get a large slice in the wheel and inactive peptides get a small slice. In the selection process one may imagine mating pairs to be selected by spinning this wheel. (Note: ALL individuals have a place on the wheel and therefore have a finite chance to be selected). This produces a list of pairs for mating. Subsequent generations are formed as follows: each member of the first generation17 is ranked by fitness, and the fittest individual is placed into the next generation with no change. Next, pairs of individuals (from the selection step above) are crossed-over to form the next generation. The crossover step may be visualized as follows (although (13) We do not use deletion and insertion as we do not want to change the overall size of the chromosome and, therefore, overall length of the bit strings. (14) Crossover is the single most important aspect which provides for most optimum assurance to explore the gene population for selecting a set of more fit members. (15) In a true sense this random generation should be referred to as generation 0 (zero) as far as GA’s are concerned, since there are no fitness functions which need to be evaluated by GA to provide the initial population of members. (16) Genesis version 1.2 from ftp site: ftp.aic.nrl.navy.mil.

Use of Genetic Algorithms in Combinatorial Synthesis some of the actual details of the implementation are slightly different for technical reasons). The genome for each individual is of fixed length (in our case 30 bits). One can envision linking up the genomes for the two members of a mating pairs and then ARBITRARILY making a “cut” at a randomly chosen spot in both of the genomes. Recombining the first part of the first genome with the second part of the second and vice versa generates two “new” offspring individuals. This entire process is called crossover step. It is important that the total number of individuals Npop selected for the subsequent generations remain identical to the initial random population, since each pair of parents produce exactly two offspring. After applying the selection and crossover steps as outlined above and thus producing a population of “new” individuals for the next generation, the mutation operator is applied. This simply consists in our case of “flipping” a bit (from 0 to 1 or vice versa). The frequency of this mutation is preset and constant throughout the run. The choice of which individual to mutate and which bit in that individuals’ genome to mutate is purely random. Chemistry. As stated earlier, we have selected a problem of protease substrate specificity and selectivity determination as a test case to evaluate the suitability and usefulness of GA in this area. A hexapeptide could be represented by the generic formula18 X1X2X3X4X5X6. Early input into the makeup of the starting population for these study was based on the previously known result that proline in position 2 of a hexapeptide increased its chance of being a substrate of the target enzyme stromelysin.11 It is normal when designing an assault on any biological target to use as much information as is available. Thus it was felt that fixing X2 ) proline for the selection of the initial “random” population set of 60 hexapeptides was a reasonable approach to biasing toward an early convergence. If this choice was not made, we reasoned that the initial population would probably all be inactive and no reasonable selection criteria could then be applied. It is however important to point out that this constraint was applied ONLY to the initial choice. Subsequent generations were free to chose non-proline amino acids for position 2. The initial random population was selected from a possible of 205 (3 200 000) possible hexapeptides (represented by X1PX3X4X5X6). However, the X2 ) Pro constraint was not imposed on subsequent generations. The peptides were synthesized using controlled-pore glass as a solid support as described previously.19 Controlled-pore glass (CPG) containing an ((aminopropyl)silyl)oxy (AMP) handle was exhaustively coupled with FMOC-β-alanine (βAla) followed by couplings with FMOC--aminocaproic acid (Acp) using a 10-fold excess of preformed HOBt active esters in Nmethylpyrrolidone. A 200 g sample of CPG containing a homogeneous population of linker [(Acp)5-βAla] was prepared. The homogeneity of FMOC-(Acp)5-βAla-CPG bulk sample was verified at each step of the coupling reaction by triplicate amino acid analysis using βAla as an internal standard. This sample was subsequently used for the preparation of the required peptides for any given generation. A standard protocol of triple coupling with a 10-fold excess of FMOC amino acid HOBt active esters (in situ activation method) was utilized for automated synthesis employing an Advanced ChemTech synthesizer model MPS 350. Finally, each peptide was reacted with FMOC alanine20a followed by capping with coumarinpropanoic acid (COP) as a fluorescent tag.20b A glass-bound peptide sample21 used for biological assays could be generically represented as COP-A-X1X2X3X4X5X6(17) Since we are using binary bit strings (0’s and 1’s), in order to represent 20 amino acids, we need 25 (i.e., 32) bits. Some of the 20 amino acids are represented by more then one string. The choice of this bit degeneracy was selected at random, but once selected, it was kept constant through out the experiment. (18) Where X represents one amino acid present at a time and the numbering is used for the sake of discussion only. (19) (a) Ator, M.; Beigel, S.; Dankanich, T.; Echols, M.; Gainor, J.; Gilliam, C.; Gordon, T.; Koch, D.; Kruse, L.; Morgan, B.; Olsen, R.; Siahaan, T.; Singh, J.; Whipple, D. Peptides: Chemistry Structure and Biology; Proceedings of the 13th American Peptide Symposium; Hodges, R., Smith, J., Eds.; 1994; pp 1012-1016. (b) We have previously described the reasons and the validation for use of CPG as a solid support, see ref 10 above for details. (20) (a) We have incorporated an alanine residue at the N-terminus of all sequences identified by GA before we tag the N-terminus with the fluorescent markersCOP. This was carried out to distance the marker group further away from the active site of a protease. (b) Gainor, J. A.; Gordon, T. D.; Morgan, B. A. Peptides: Chemistry Structure and Biology; Proceedings of the 13th American Peptide Symposium; Hodges, R., Smith, J., Eds.; 1994; pp 989-991.

J. Am. Chem. Soc., Vol. 118, No. 7, 1996 1671 (Acp)5-βAla-AMP-CPG, where Xi (i ) 1-6) represents one of the 20 possible amino acids (see Abbreviations for the list of amino acids used). Biological Assays. Automated biological assays were performed on a small sample of the glass-bound peptides in a 96-well format. Typically, 4.0 ( 0.3 mg of 62 glass-bound peptides were weighed22 in individual tubes using the HP ORCA robot and 220 µL of a buffer (containing 50 mM Tris (pH 7.5), 200 mM NaCl, and 10 mM CaCl2) followed by 44 µL of the protease solution (20 nM mSl-t or 10 nM mCl-t)23 was added using a Packard PROBE. Samples were mixed on a variable speed vortexer for 120 min. The substrate was allowed to settle by gravity, a 125 µL aliquot of each sample was transferred to the appropriate position of a 96-well plate, and fluorescence was read using a fluorescence plate reader. The stromelysin construct utilized has some autocatalytic activity. The positive control sample (GPLAMF) and a negative control sample (hexa-D-alanyl) were synthesized and evaluated as part of each generation. The assay results for these samples were used as an indication of the validity of the assay and to account for variance in stromelysin activity due to autocatalysis. The negative control samples typically produced very low fluorescence (